Fall Semester Announcements

What does my future hold?

Update: I haven’t joined Anthropic or taken leave from the university.

That means I’ll be teaching as usual this fall, and you can look forward to your regular installment of live lecture blogs. Unlike past years, I’m assigned to my grad seminar in the fall, not the spring. Next semester I’ll be teaching our undergraduate probability class. That is assuming that the University is still here in the spring and hasn’t been put out of business by some new AI super tutor released by my friends across the bay. I mean, it would be the year 2027, and popular forecasts suggest AIs will be able to do everything taught in a CS degree by late 2026. By the end of the spring semester, those same prognosticators predict those AIs will go rogue, and we’ll find ourselves in the reality forecast by James Cameron in his 1984 prophecy, Terminator. Why should I bother dusting off my copy of Bertsekas and Tsitsiklis when the forecasts tell me I should work harder at the gym to prepare myself for robot enslavement in the salt mines?

Actually, you know what would be a good way to prep for the robot apocalypse? Why don’t we spend a semester talking about forecasting and why people are obsessed with being certain about the future?

Answering that question could potentially be a great way to shape a graduate course. We could spend a semester digging into not only how people forecast, but why they forecast. We could split our time in half, looking at the particularities of different domains where people make forecasts, and then looking into the tools they have settled on as mathematical culture.

If you look at the places where forecasts are common, they all have different purposes. A local weather report is very different from a prediction of the end of the world. The former tells you if you should pack an umbrella on the way to work. The latter tells you whether you need to lobby your government to radically change its planned energy buildout. We all believe the evidence supporting forecasts of rain is far more certain and reliable than forecasts of nature’s end. But the costs of being wrong couldn’t be more different.

What impact do these forecasts have? Why do we forecast the weather, and what hidden technology is needed to make these forecasts accurate? Why does Congress demand that we forecast future budgetary consequences of proposed laws, even though we know we can’t predict the actual structural shocks that will render those forecasts moot? Why are people so obsessed with predicting the rise of superintelligent robots? Is it more than a way to justify their greed and obsessive 996 work conditions?

We’d learn a lot from a comparative study. I’d like to look at astronomy, meteorology, climate science, seismology, macroeconomics, government, epidemiology, public opinion research, and millenarianism to see what they have in common and how they differ. Forecasts let people externalize their beliefs about the likelihoods and consequences of various scenarios. Some forecasts are made for mundane planning. Some forecasts are made to literally gamble. Some forecasts communicate possible futures that others might not be considering. Some forecasts are made to be self-fulfilling, to manifest a change in the world the forecaster desires. Some are made to be self-negating, to convince people to act to avoid worst-case scenarios. I’m interested in understanding the threads that link all of these different purposes together.

Though I’m much more interested in the why, the how has some fun tidbits too. The how is about creating certainty about the future by quantifying it. Uncertainty becomes certain once we turn it into an interval, right? In talking about the how-of-forecasting, I could tie together a lot of loose ends in methods that I’ve blogged about over the years.

We might determine the conditions under which pattern recognition (aka machine learning aka AI) becomes a forecast. We could look at how forecasts are evaluated post-hoc with scoring rules and calibration tests and why people think those are good evaluations. We could look at methods for cost-benefit analysis and uncertainty quantification, and how people justify their modeling assumptions to make decisions. We could learn about tools from dynamical systems that move from simple moving averages to complex simulations. We could examine how statistical tools can be applied to extrapolate from the present to the future. And we could see how these sorts of metrics and models tie your hands algorithmically into unsurprising answers.

This sounds like a fun class to me. I predict I’ll teach this class this fall and live blog it here, starting this Thursday. If you’re a Berkeley graduate student whose research depends on forecasts, email me if you’d like to join the course.[footnote: If you do email, please send me a description of your background and why you’re interested.] If you’re not local, I’ll post a syllabus and webpage this week, and I’ll do my best to keep all of the material public. I predict it will be fun.